linux/fs/direct-io.c

1229 lines
34 KiB
C
Raw Normal View History

/*
* fs/direct-io.c
*
* Copyright (C) 2002, Linus Torvalds.
*
* O_DIRECT
*
* 04Jul2002 akpm@zip.com.au
* Initial version
* 11Sep2002 janetinc@us.ibm.com
* added readv/writev support.
* 29Oct2002 akpm@zip.com.au
* rewrote bio_add_page() support.
* 30Oct2002 pbadari@us.ibm.com
* added support for non-aligned IO.
* 06Nov2002 pbadari@us.ibm.com
* added asynchronous IO support.
* 21Jul2003 nathans@sgi.com
* added IO completion notifier.
*/
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/types.h>
#include <linux/fs.h>
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/pagemap.h>
#include <linux/task_io_accounting_ops.h>
#include <linux/bio.h>
#include <linux/wait.h>
#include <linux/err.h>
#include <linux/blkdev.h>
#include <linux/buffer_head.h>
#include <linux/rwsem.h>
#include <linux/uio.h>
#include <asm/atomic.h>
/*
* How many user pages to map in one call to get_user_pages(). This determines
* the size of a structure on the stack.
*/
#define DIO_PAGES 64
/*
* This code generally works in units of "dio_blocks". A dio_block is
* somewhere between the hard sector size and the filesystem block size. it
* is determined on a per-invocation basis. When talking to the filesystem
* we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
* down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
* to bio_block quantities by shifting left by blkfactor.
*
* If blkfactor is zero then the user's request was aligned to the filesystem's
* blocksize.
*
* lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
* This determines whether we need to do the fancy locking which prevents
* direct-IO from being able to read uninitialised disk blocks. If its zero
* (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
* not held for the entire direct write (taken briefly, initially, during a
* direct read though, but its never held for the duration of a direct-IO).
*/
struct dio {
/* BIO submission state */
struct bio *bio; /* bio under assembly */
struct inode *inode;
int rw;
loff_t i_size; /* i_size when submitted */
int lock_type; /* doesn't change */
unsigned blkbits; /* doesn't change */
unsigned blkfactor; /* When we're using an alignment which
is finer than the filesystem's soft
blocksize, this specifies how much
finer. blkfactor=2 means 1/4-block
alignment. Does not change */
unsigned start_zero_done; /* flag: sub-blocksize zeroing has
been performed at the start of a
write */
int pages_in_io; /* approximate total IO pages */
size_t size; /* total request size (doesn't change)*/
sector_t block_in_file; /* Current offset into the underlying
file in dio_block units. */
unsigned blocks_available; /* At block_in_file. changes */
sector_t final_block_in_request;/* doesn't change */
unsigned first_block_in_page; /* doesn't change, Used only once */
int boundary; /* prev block is at a boundary */
int reap_counter; /* rate limit reaping */
get_block_t *get_block; /* block mapping function */
dio_iodone_t *end_io; /* IO completion function */
sector_t final_block_in_bio; /* current final block in bio + 1 */
sector_t next_block_for_io; /* next block to be put under IO,
in dio_blocks units */
struct buffer_head map_bh; /* last get_block() result */
/*
* Deferred addition of a page to the dio. These variables are
* private to dio_send_cur_page(), submit_page_section() and
* dio_bio_add_page().
*/
struct page *cur_page; /* The page */
unsigned cur_page_offset; /* Offset into it, in bytes */
unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
sector_t cur_page_block; /* Where it starts */
/*
* Page fetching state. These variables belong to dio_refill_pages().
*/
int curr_page; /* changes */
int total_pages; /* doesn't change */
unsigned long curr_user_address;/* changes */
/*
* Page queue. These variables belong to dio_refill_pages() and
* dio_get_page().
*/
struct page *pages[DIO_PAGES]; /* page buffer */
unsigned head; /* next page to process */
unsigned tail; /* last valid page + 1 */
int page_errors; /* errno from get_user_pages() */
/* BIO completion state */
spinlock_t bio_lock; /* protects BIO fields below */
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
unsigned long refcount; /* direct_io_worker() and bios */
struct bio *bio_list; /* singly linked via bi_private */
struct task_struct *waiter; /* waiting task (NULL if none) */
/* AIO related stuff */
struct kiocb *iocb; /* kiocb */
int is_async; /* is IO async ? */
[PATCH] direct-io: bug fix in dio handling write error There is a bug in direct-io on propagating write error up to the higher I/O layer. When performing an async ODIRECT write to a block device, if a device error occurred (like media error or disk is pulled), the error code is only propagated from device driver to the DIO layer. The error code stops at finished_one_bio(). The aysnc write, however, is supposedly have a corresponding AIO event with appropriate return code (in this case -EIO). Application which waits on the async write event, will hang forever since such AIO event is lost forever (if such app did not use the timeout option in io_getevents call. Regardless, an AIO event is lost). The discovery of above bug leads to another discovery of potential race window with dio->result. The fundamental problem is that dio->result is overloaded with dual use: an indicator of fall back path for partial dio write, and an error indicator used in the I/O completion path. In the event of device error, the setting of -EIO to dio->result clashes with value used to track partial write that activates the fall back path. It was also pointed out that it is impossible to use dio->result to track partial write and at the same time to track error returned from device driver. Because direct_io_work can only determines whether it is a partial write at the end of io submission and in mid stream of those io submission, a return code could be coming back from the driver. Thus messing up all the subsequent logic. Proposed fix is to separating out error code returned by the IO completion path from partial IO submit tracking. A new variable is added to dio structure specifically to track io error returned in the completion path. Signed-off-by: Ken Chen <kenneth.w.chen@intel.com> Acked-by: Zach Brown <zach.brown@oracle.com> Acked-by: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 19:08:16 +08:00
int io_error; /* IO error in completion path */
ssize_t result; /* IO result */
};
/*
* How many pages are in the queue?
*/
static inline unsigned dio_pages_present(struct dio *dio)
{
return dio->tail - dio->head;
}
/*
* Go grab and pin some userspace pages. Typically we'll get 64 at a time.
*/
static int dio_refill_pages(struct dio *dio)
{
int ret;
int nr_pages;
nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
down_read(&current->mm->mmap_sem);
ret = get_user_pages(
current, /* Task for fault acounting */
current->mm, /* whose pages? */
dio->curr_user_address, /* Where from? */
nr_pages, /* How many pages? */
dio->rw == READ, /* Write to memory? */
0, /* force (?) */
&dio->pages[0],
NULL); /* vmas */
up_read(&current->mm->mmap_sem);
if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
remove ZERO_PAGE The commit b5810039a54e5babf428e9a1e89fc1940fabff11 contains the note A last caveat: the ZERO_PAGE is now refcounted and managed with rmap (and thus mapcounted and count towards shared rss). These writes to the struct page could cause excessive cacheline bouncing on big systems. There are a number of ways this could be addressed if it is an issue. And indeed this cacheline bouncing has shown up on large SGI systems. There was a situation where an Altix system was essentially livelocked tearing down ZERO_PAGE pagetables when an HPC app aborted during startup. This situation can be avoided in userspace, but it does highlight the potential scalability problem with refcounting ZERO_PAGE, and corner cases where it can really hurt (we don't want the system to livelock!). There are several broad ways to fix this problem: 1. add back some special casing to avoid refcounting ZERO_PAGE 2. per-node or per-cpu ZERO_PAGES 3. remove the ZERO_PAGE completely I will argue for 3. The others should also fix the problem, but they result in more complex code than does 3, with little or no real benefit that I can see. Why? Inserting a ZERO_PAGE for anonymous read faults appears to be a false optimisation: if an application is performance critical, it would not be doing many read faults of new memory, or at least it could be expected to write to that memory soon afterwards. If cache or memory use is critical, it should not be working with a significant number of ZERO_PAGEs anyway (a more compact representation of zeroes should be used). As a sanity check -- mesuring on my desktop system, there are never many mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not increase much without it. When running a make -j4 kernel compile on my dual core system, there are about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000 ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second is torn down without being COWed). So removing ZERO_PAGE will save 1,000 page faults per second when running kbuild, while keeping it only saves less than 1 page clearing operation per second. 1 page clear is cheaper than a thousand faults, presumably, so there isn't an obvious loss. Neither the logical argument nor these basic tests give a guarantee of no regressions. However, this is a reasonable opportunity to try to remove the ZERO_PAGE from the pagefault path. If it is found to cause regressions, we can reintroduce it and just avoid refcounting it. The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked. I don't see much use to them except on benchmarks. All other users of ZERO_PAGE are converted just to use ZERO_PAGE(0) for simplicity. We can look at replacing them all and maybe ripping out ZERO_PAGE completely when we are more satisfied with this solution. Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus "snif" Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:40 +08:00
struct page *page = ZERO_PAGE(0);
/*
* A memory fault, but the filesystem has some outstanding
* mapped blocks. We need to use those blocks up to avoid
* leaking stale data in the file.
*/
if (dio->page_errors == 0)
dio->page_errors = ret;
2005-10-30 09:16:12 +08:00
page_cache_get(page);
dio->pages[0] = page;
dio->head = 0;
dio->tail = 1;
ret = 0;
goto out;
}
if (ret >= 0) {
dio->curr_user_address += ret * PAGE_SIZE;
dio->curr_page += ret;
dio->head = 0;
dio->tail = ret;
ret = 0;
}
out:
return ret;
}
/*
* Get another userspace page. Returns an ERR_PTR on error. Pages are
* buffered inside the dio so that we can call get_user_pages() against a
* decent number of pages, less frequently. To provide nicer use of the
* L1 cache.
*/
static struct page *dio_get_page(struct dio *dio)
{
if (dio_pages_present(dio) == 0) {
int ret;
ret = dio_refill_pages(dio);
if (ret)
return ERR_PTR(ret);
BUG_ON(dio_pages_present(dio) == 0);
}
return dio->pages[dio->head++];
}
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
/**
* dio_complete() - called when all DIO BIO I/O has been completed
* @offset: the byte offset in the file of the completed operation
*
* This releases locks as dictated by the locking type, lets interested parties
* know that a DIO operation has completed, and calculates the resulting return
* code for the operation.
*
* It lets the filesystem know if it registered an interest earlier via
* get_block. Pass the private field of the map buffer_head so that
* filesystems can use it to hold additional state between get_block calls and
* dio_complete.
*/
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
static int dio_complete(struct dio *dio, loff_t offset, int ret)
{
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
ssize_t transferred = 0;
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
/*
* AIO submission can race with bio completion to get here while
* expecting to have the last io completed by bio completion.
* In that case -EIOCBQUEUED is in fact not an error we want
* to preserve through this call.
*/
if (ret == -EIOCBQUEUED)
ret = 0;
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
if (dio->result) {
transferred = dio->result;
/* Check for short read case */
if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
transferred = dio->i_size - offset;
}
if (dio->end_io && dio->result)
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
dio->end_io(dio->iocb, offset, transferred,
dio->map_bh.b_private);
if (dio->lock_type == DIO_LOCKING)
/* lockdep: non-owner release */
up_read_non_owner(&dio->inode->i_alloc_sem);
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
if (ret == 0)
ret = dio->page_errors;
if (ret == 0)
ret = dio->io_error;
if (ret == 0)
ret = transferred;
return ret;
}
static int dio_bio_complete(struct dio *dio, struct bio *bio);
/*
* Asynchronous IO callback.
*/
static void dio_bio_end_aio(struct bio *bio, int error)
{
struct dio *dio = bio->bi_private;
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
unsigned long remaining;
unsigned long flags;
/* cleanup the bio */
dio_bio_complete(dio, bio);
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
spin_lock_irqsave(&dio->bio_lock, flags);
remaining = --dio->refcount;
if (remaining == 1 && dio->waiter)
wake_up_process(dio->waiter);
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
spin_unlock_irqrestore(&dio->bio_lock, flags);
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
if (remaining == 0) {
int ret = dio_complete(dio, dio->iocb->ki_pos, 0);
aio_complete(dio->iocb, ret, 0);
kfree(dio);
}
}
/*
* The BIO completion handler simply queues the BIO up for the process-context
* handler.
*
* During I/O bi_private points at the dio. After I/O, bi_private is used to
* implement a singly-linked list of completed BIOs, at dio->bio_list.
*/
static void dio_bio_end_io(struct bio *bio, int error)
{
struct dio *dio = bio->bi_private;
unsigned long flags;
spin_lock_irqsave(&dio->bio_lock, flags);
bio->bi_private = dio->bio_list;
dio->bio_list = bio;
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
if (--dio->refcount == 1 && dio->waiter)
wake_up_process(dio->waiter);
spin_unlock_irqrestore(&dio->bio_lock, flags);
}
static int
dio_bio_alloc(struct dio *dio, struct block_device *bdev,
sector_t first_sector, int nr_vecs)
{
struct bio *bio;
bio = bio_alloc(GFP_KERNEL, nr_vecs);
if (bio == NULL)
return -ENOMEM;
bio->bi_bdev = bdev;
bio->bi_sector = first_sector;
if (dio->is_async)
bio->bi_end_io = dio_bio_end_aio;
else
bio->bi_end_io = dio_bio_end_io;
dio->bio = bio;
return 0;
}
/*
* In the AIO read case we speculatively dirty the pages before starting IO.
* During IO completion, any of these pages which happen to have been written
* back will be redirtied by bio_check_pages_dirty().
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
*
* bios hold a dio reference between submit_bio and ->end_io.
*/
static void dio_bio_submit(struct dio *dio)
{
struct bio *bio = dio->bio;
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
unsigned long flags;
bio->bi_private = dio;
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
spin_lock_irqsave(&dio->bio_lock, flags);
dio->refcount++;
spin_unlock_irqrestore(&dio->bio_lock, flags);
if (dio->is_async && dio->rw == READ)
bio_set_pages_dirty(bio);
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
submit_bio(dio->rw, bio);
dio->bio = NULL;
dio->boundary = 0;
}
/*
* Release any resources in case of a failure
*/
static void dio_cleanup(struct dio *dio)
{
while (dio_pages_present(dio))
page_cache_release(dio_get_page(dio));
}
/*
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
* Wait for the next BIO to complete. Remove it and return it. NULL is
* returned once all BIOs have been completed. This must only be called once
* all bios have been issued so that dio->refcount can only decrease. This
* requires that that the caller hold a reference on the dio.
*/
static struct bio *dio_await_one(struct dio *dio)
{
unsigned long flags;
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
struct bio *bio = NULL;
spin_lock_irqsave(&dio->bio_lock, flags);
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
/*
* Wait as long as the list is empty and there are bios in flight. bio
* completion drops the count, maybe adds to the list, and wakes while
* holding the bio_lock so we don't need set_current_state()'s barrier
* and can call it after testing our condition.
*/
while (dio->refcount > 1 && dio->bio_list == NULL) {
__set_current_state(TASK_UNINTERRUPTIBLE);
dio->waiter = current;
spin_unlock_irqrestore(&dio->bio_lock, flags);
io_schedule();
/* wake up sets us TASK_RUNNING */
spin_lock_irqsave(&dio->bio_lock, flags);
dio->waiter = NULL;
}
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
if (dio->bio_list) {
bio = dio->bio_list;
dio->bio_list = bio->bi_private;
}
spin_unlock_irqrestore(&dio->bio_lock, flags);
return bio;
}
/*
* Process one completed BIO. No locks are held.
*/
static int dio_bio_complete(struct dio *dio, struct bio *bio)
{
const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
struct bio_vec *bvec = bio->bi_io_vec;
int page_no;
if (!uptodate)
[PATCH] direct-io: bug fix in dio handling write error There is a bug in direct-io on propagating write error up to the higher I/O layer. When performing an async ODIRECT write to a block device, if a device error occurred (like media error or disk is pulled), the error code is only propagated from device driver to the DIO layer. The error code stops at finished_one_bio(). The aysnc write, however, is supposedly have a corresponding AIO event with appropriate return code (in this case -EIO). Application which waits on the async write event, will hang forever since such AIO event is lost forever (if such app did not use the timeout option in io_getevents call. Regardless, an AIO event is lost). The discovery of above bug leads to another discovery of potential race window with dio->result. The fundamental problem is that dio->result is overloaded with dual use: an indicator of fall back path for partial dio write, and an error indicator used in the I/O completion path. In the event of device error, the setting of -EIO to dio->result clashes with value used to track partial write that activates the fall back path. It was also pointed out that it is impossible to use dio->result to track partial write and at the same time to track error returned from device driver. Because direct_io_work can only determines whether it is a partial write at the end of io submission and in mid stream of those io submission, a return code could be coming back from the driver. Thus messing up all the subsequent logic. Proposed fix is to separating out error code returned by the IO completion path from partial IO submit tracking. A new variable is added to dio structure specifically to track io error returned in the completion path. Signed-off-by: Ken Chen <kenneth.w.chen@intel.com> Acked-by: Zach Brown <zach.brown@oracle.com> Acked-by: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-03-25 19:08:16 +08:00
dio->io_error = -EIO;
if (dio->is_async && dio->rw == READ) {
bio_check_pages_dirty(bio); /* transfers ownership */
} else {
for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
struct page *page = bvec[page_no].bv_page;
if (dio->rw == READ && !PageCompound(page))
set_page_dirty_lock(page);
page_cache_release(page);
}
bio_put(bio);
}
return uptodate ? 0 : -EIO;
}
/*
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
* Wait on and process all in-flight BIOs. This must only be called once
* all bios have been issued so that the refcount can only decrease.
* This just waits for all bios to make it through dio_bio_complete. IO
* errors are propagated through dio->io_error and should be propagated via
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
* dio_complete().
*/
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
static void dio_await_completion(struct dio *dio)
{
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
struct bio *bio;
do {
bio = dio_await_one(dio);
if (bio)
dio_bio_complete(dio, bio);
} while (bio);
}
/*
* A really large O_DIRECT read or write can generate a lot of BIOs. So
* to keep the memory consumption sane we periodically reap any completed BIOs
* during the BIO generation phase.
*
* This also helps to limit the peak amount of pinned userspace memory.
*/
static int dio_bio_reap(struct dio *dio)
{
int ret = 0;
if (dio->reap_counter++ >= 64) {
while (dio->bio_list) {
unsigned long flags;
struct bio *bio;
int ret2;
spin_lock_irqsave(&dio->bio_lock, flags);
bio = dio->bio_list;
dio->bio_list = bio->bi_private;
spin_unlock_irqrestore(&dio->bio_lock, flags);
ret2 = dio_bio_complete(dio, bio);
if (ret == 0)
ret = ret2;
}
dio->reap_counter = 0;
}
return ret;
}
/*
* Call into the fs to map some more disk blocks. We record the current number
* of available blocks at dio->blocks_available. These are in units of the
* fs blocksize, (1 << inode->i_blkbits).
*
* The fs is allowed to map lots of blocks at once. If it wants to do that,
* it uses the passed inode-relative block number as the file offset, as usual.
*
* get_block() is passed the number of i_blkbits-sized blocks which direct_io
* has remaining to do. The fs should not map more than this number of blocks.
*
* If the fs has mapped a lot of blocks, it should populate bh->b_size to
* indicate how much contiguous disk space has been made available at
* bh->b_blocknr.
*
* If *any* of the mapped blocks are new, then the fs must set buffer_new().
* This isn't very efficient...
*
* In the case of filesystem holes: the fs may return an arbitrarily-large
* hole by returning an appropriate value in b_size and by clearing
* buffer_mapped(). However the direct-io code will only process holes one
* block at a time - it will repeatedly call get_block() as it walks the hole.
*/
static int get_more_blocks(struct dio *dio)
{
int ret;
struct buffer_head *map_bh = &dio->map_bh;
sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
unsigned long fs_count; /* Number of filesystem-sized blocks */
unsigned long dio_count;/* Number of dio_block-sized blocks */
unsigned long blkmask;
int create;
/*
* If there was a memory error and we've overwritten all the
* mapped blocks then we can now return that memory error
*/
ret = dio->page_errors;
if (ret == 0) {
BUG_ON(dio->block_in_file >= dio->final_block_in_request);
fs_startblk = dio->block_in_file >> dio->blkfactor;
dio_count = dio->final_block_in_request - dio->block_in_file;
fs_count = dio_count >> dio->blkfactor;
blkmask = (1 << dio->blkfactor) - 1;
if (dio_count & blkmask)
fs_count++;
map_bh->b_state = 0;
map_bh->b_size = fs_count << dio->inode->i_blkbits;
create = dio->rw & WRITE;
if (dio->lock_type == DIO_LOCKING) {
if (dio->block_in_file < (i_size_read(dio->inode) >>
dio->blkbits))
create = 0;
} else if (dio->lock_type == DIO_NO_LOCKING) {
create = 0;
}
/*
* For writes inside i_size we forbid block creations: only
* overwrites are permitted. We fall back to buffered writes
* at a higher level for inside-i_size block-instantiating
* writes.
*/
ret = (*dio->get_block)(dio->inode, fs_startblk,
map_bh, create);
}
return ret;
}
/*
* There is no bio. Make one now.
*/
static int dio_new_bio(struct dio *dio, sector_t start_sector)
{
sector_t sector;
int ret, nr_pages;
ret = dio_bio_reap(dio);
if (ret)
goto out;
sector = start_sector << (dio->blkbits - 9);
nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
BUG_ON(nr_pages <= 0);
ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
dio->boundary = 0;
out:
return ret;
}
/*
* Attempt to put the current chunk of 'cur_page' into the current BIO. If
* that was successful then update final_block_in_bio and take a ref against
* the just-added page.
*
* Return zero on success. Non-zero means the caller needs to start a new BIO.
*/
static int dio_bio_add_page(struct dio *dio)
{
int ret;
ret = bio_add_page(dio->bio, dio->cur_page,
dio->cur_page_len, dio->cur_page_offset);
if (ret == dio->cur_page_len) {
/*
* Decrement count only, if we are done with this page
*/
if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
dio->pages_in_io--;
page_cache_get(dio->cur_page);
dio->final_block_in_bio = dio->cur_page_block +
(dio->cur_page_len >> dio->blkbits);
ret = 0;
} else {
ret = 1;
}
return ret;
}
/*
* Put cur_page under IO. The section of cur_page which is described by
* cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
* starts on-disk at cur_page_block.
*
* We take a ref against the page here (on behalf of its presence in the bio).
*
* The caller of this function is responsible for removing cur_page from the
* dio, and for dropping the refcount which came from that presence.
*/
static int dio_send_cur_page(struct dio *dio)
{
int ret = 0;
if (dio->bio) {
/*
* See whether this new request is contiguous with the old
*/
if (dio->final_block_in_bio != dio->cur_page_block)
dio_bio_submit(dio);
/*
* Submit now if the underlying fs is about to perform a
* metadata read
*/
if (dio->boundary)
dio_bio_submit(dio);
}
if (dio->bio == NULL) {
ret = dio_new_bio(dio, dio->cur_page_block);
if (ret)
goto out;
}
if (dio_bio_add_page(dio) != 0) {
dio_bio_submit(dio);
ret = dio_new_bio(dio, dio->cur_page_block);
if (ret == 0) {
ret = dio_bio_add_page(dio);
BUG_ON(ret != 0);
}
}
out:
return ret;
}
/*
* An autonomous function to put a chunk of a page under deferred IO.
*
* The caller doesn't actually know (or care) whether this piece of page is in
* a BIO, or is under IO or whatever. We just take care of all possible
* situations here. The separation between the logic of do_direct_IO() and
* that of submit_page_section() is important for clarity. Please don't break.
*
* The chunk of page starts on-disk at blocknr.
*
* We perform deferred IO, by recording the last-submitted page inside our
* private part of the dio structure. If possible, we just expand the IO
* across that page here.
*
* If that doesn't work out then we put the old page into the bio and add this
* page to the dio instead.
*/
static int
submit_page_section(struct dio *dio, struct page *page,
unsigned offset, unsigned len, sector_t blocknr)
{
int ret = 0;
if (dio->rw & WRITE) {
/*
* Read accounting is performed in submit_bio()
*/
task_io_account_write(len);
}
/*
* Can we just grow the current page's presence in the dio?
*/
if ( (dio->cur_page == page) &&
(dio->cur_page_offset + dio->cur_page_len == offset) &&
(dio->cur_page_block +
(dio->cur_page_len >> dio->blkbits) == blocknr)) {
dio->cur_page_len += len;
/*
* If dio->boundary then we want to schedule the IO now to
* avoid metadata seeks.
*/
if (dio->boundary) {
ret = dio_send_cur_page(dio);
page_cache_release(dio->cur_page);
dio->cur_page = NULL;
}
goto out;
}
/*
* If there's a deferred page already there then send it.
*/
if (dio->cur_page) {
ret = dio_send_cur_page(dio);
page_cache_release(dio->cur_page);
dio->cur_page = NULL;
if (ret)
goto out;
}
page_cache_get(page); /* It is in dio */
dio->cur_page = page;
dio->cur_page_offset = offset;
dio->cur_page_len = len;
dio->cur_page_block = blocknr;
out:
return ret;
}
/*
* Clean any dirty buffers in the blockdev mapping which alias newly-created
* file blocks. Only called for S_ISREG files - blockdevs do not set
* buffer_new
*/
static void clean_blockdev_aliases(struct dio *dio)
{
unsigned i;
unsigned nblocks;
nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
for (i = 0; i < nblocks; i++) {
unmap_underlying_metadata(dio->map_bh.b_bdev,
dio->map_bh.b_blocknr + i);
}
}
/*
* If we are not writing the entire block and get_block() allocated
* the block for us, we need to fill-in the unused portion of the
* block with zeros. This happens only if user-buffer, fileoffset or
* io length is not filesystem block-size multiple.
*
* `end' is zero if we're doing the start of the IO, 1 at the end of the
* IO.
*/
static void dio_zero_block(struct dio *dio, int end)
{
unsigned dio_blocks_per_fs_block;
unsigned this_chunk_blocks; /* In dio_blocks */
unsigned this_chunk_bytes;
struct page *page;
dio->start_zero_done = 1;
if (!dio->blkfactor || !buffer_new(&dio->map_bh))
return;
dio_blocks_per_fs_block = 1 << dio->blkfactor;
this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
if (!this_chunk_blocks)
return;
/*
* We need to zero out part of an fs block. It is either at the
* beginning or the end of the fs block.
*/
if (end)
this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
this_chunk_bytes = this_chunk_blocks << dio->blkbits;
remove ZERO_PAGE The commit b5810039a54e5babf428e9a1e89fc1940fabff11 contains the note A last caveat: the ZERO_PAGE is now refcounted and managed with rmap (and thus mapcounted and count towards shared rss). These writes to the struct page could cause excessive cacheline bouncing on big systems. There are a number of ways this could be addressed if it is an issue. And indeed this cacheline bouncing has shown up on large SGI systems. There was a situation where an Altix system was essentially livelocked tearing down ZERO_PAGE pagetables when an HPC app aborted during startup. This situation can be avoided in userspace, but it does highlight the potential scalability problem with refcounting ZERO_PAGE, and corner cases where it can really hurt (we don't want the system to livelock!). There are several broad ways to fix this problem: 1. add back some special casing to avoid refcounting ZERO_PAGE 2. per-node or per-cpu ZERO_PAGES 3. remove the ZERO_PAGE completely I will argue for 3. The others should also fix the problem, but they result in more complex code than does 3, with little or no real benefit that I can see. Why? Inserting a ZERO_PAGE for anonymous read faults appears to be a false optimisation: if an application is performance critical, it would not be doing many read faults of new memory, or at least it could be expected to write to that memory soon afterwards. If cache or memory use is critical, it should not be working with a significant number of ZERO_PAGEs anyway (a more compact representation of zeroes should be used). As a sanity check -- mesuring on my desktop system, there are never many mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not increase much without it. When running a make -j4 kernel compile on my dual core system, there are about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000 ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second is torn down without being COWed). So removing ZERO_PAGE will save 1,000 page faults per second when running kbuild, while keeping it only saves less than 1 page clearing operation per second. 1 page clear is cheaper than a thousand faults, presumably, so there isn't an obvious loss. Neither the logical argument nor these basic tests give a guarantee of no regressions. However, this is a reasonable opportunity to try to remove the ZERO_PAGE from the pagefault path. If it is found to cause regressions, we can reintroduce it and just avoid refcounting it. The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked. I don't see much use to them except on benchmarks. All other users of ZERO_PAGE are converted just to use ZERO_PAGE(0) for simplicity. We can look at replacing them all and maybe ripping out ZERO_PAGE completely when we are more satisfied with this solution. Signed-off-by: Nick Piggin <npiggin@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus "snif" Torvalds <torvalds@linux-foundation.org>
2007-10-16 16:24:40 +08:00
page = ZERO_PAGE(0);
if (submit_page_section(dio, page, 0, this_chunk_bytes,
dio->next_block_for_io))
return;
dio->next_block_for_io += this_chunk_blocks;
}
/*
* Walk the user pages, and the file, mapping blocks to disk and generating
* a sequence of (page,offset,len,block) mappings. These mappings are injected
* into submit_page_section(), which takes care of the next stage of submission
*
* Direct IO against a blockdev is different from a file. Because we can
* happily perform page-sized but 512-byte aligned IOs. It is important that
* blockdev IO be able to have fine alignment and large sizes.
*
* So what we do is to permit the ->get_block function to populate bh.b_size
* with the size of IO which is permitted at this offset and this i_blkbits.
*
* For best results, the blockdev should be set up with 512-byte i_blkbits and
* it should set b_size to PAGE_SIZE or more inside get_block(). This gives
* fine alignment but still allows this function to work in PAGE_SIZE units.
*/
static int do_direct_IO(struct dio *dio)
{
const unsigned blkbits = dio->blkbits;
const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
struct page *page;
unsigned block_in_page;
struct buffer_head *map_bh = &dio->map_bh;
int ret = 0;
/* The I/O can start at any block offset within the first page */
block_in_page = dio->first_block_in_page;
while (dio->block_in_file < dio->final_block_in_request) {
page = dio_get_page(dio);
if (IS_ERR(page)) {
ret = PTR_ERR(page);
goto out;
}
while (block_in_page < blocks_per_page) {
unsigned offset_in_page = block_in_page << blkbits;
unsigned this_chunk_bytes; /* # of bytes mapped */
unsigned this_chunk_blocks; /* # of blocks */
unsigned u;
if (dio->blocks_available == 0) {
/*
* Need to go and map some more disk
*/
unsigned long blkmask;
unsigned long dio_remainder;
ret = get_more_blocks(dio);
if (ret) {
page_cache_release(page);
goto out;
}
if (!buffer_mapped(map_bh))
goto do_holes;
dio->blocks_available =
map_bh->b_size >> dio->blkbits;
dio->next_block_for_io =
map_bh->b_blocknr << dio->blkfactor;
if (buffer_new(map_bh))
clean_blockdev_aliases(dio);
if (!dio->blkfactor)
goto do_holes;
blkmask = (1 << dio->blkfactor) - 1;
dio_remainder = (dio->block_in_file & blkmask);
/*
* If we are at the start of IO and that IO
* starts partway into a fs-block,
* dio_remainder will be non-zero. If the IO
* is a read then we can simply advance the IO
* cursor to the first block which is to be
* read. But if the IO is a write and the
* block was newly allocated we cannot do that;
* the start of the fs block must be zeroed out
* on-disk
*/
if (!buffer_new(map_bh))
dio->next_block_for_io += dio_remainder;
dio->blocks_available -= dio_remainder;
}
do_holes:
/* Handle holes */
if (!buffer_mapped(map_bh)) {
loff_t i_size_aligned;
/* AKPM: eargh, -ENOTBLK is a hack */
if (dio->rw & WRITE) {
page_cache_release(page);
return -ENOTBLK;
}
/*
* Be sure to account for a partial block as the
* last block in the file
*/
i_size_aligned = ALIGN(i_size_read(dio->inode),
1 << blkbits);
if (dio->block_in_file >=
i_size_aligned >> blkbits) {
/* We hit eof */
page_cache_release(page);
goto out;
}
Pagecache zeroing: zero_user_segment, zero_user_segments and zero_user Simplify page cache zeroing of segments of pages through 3 functions zero_user_segments(page, start1, end1, start2, end2) Zeros two segments of the page. It takes the position where to start and end the zeroing which avoids length calculations and makes code clearer. zero_user_segment(page, start, end) Same for a single segment. zero_user(page, start, length) Length variant for the case where we know the length. We remove the zero_user_page macro. Issues: 1. Its a macro. Inline functions are preferable. 2. The KM_USER0 macro is only defined for HIGHMEM. Having to treat this special case everywhere makes the code needlessly complex. The parameter for zeroing is always KM_USER0 except in one single case that we open code. Avoiding KM_USER0 makes a lot of code not having to be dealing with the special casing for HIGHMEM anymore. Dealing with kmap is only necessary for HIGHMEM configurations. In those configurations we use KM_USER0 like we do for a series of other functions defined in highmem.h. Since KM_USER0 is depends on HIGHMEM the existing zero_user_page function could not be a macro. zero_user_* functions introduced here can be be inline because that constant is not used when these functions are called. Also extract the flushing of the caches to be outside of the kmap. [akpm@linux-foundation.org: fix nfs and ntfs build] [akpm@linux-foundation.org: fix ntfs build some more] Signed-off-by: Christoph Lameter <clameter@sgi.com> Cc: Steven French <sfrench@us.ibm.com> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Cc: <linux-ext4@vger.kernel.org> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: "J. Bruce Fields" <bfields@fieldses.org> Cc: Anton Altaparmakov <aia21@cantab.net> Cc: Mark Fasheh <mark.fasheh@oracle.com> Cc: David Chinner <dgc@sgi.com> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Cc: Steven French <sfrench@us.ibm.com> Cc: Steven Whitehouse <swhiteho@redhat.com> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-02-05 14:28:29 +08:00
zero_user(page, block_in_page << blkbits,
1 << blkbits);
dio->block_in_file++;
block_in_page++;
goto next_block;
}
/*
* If we're performing IO which has an alignment which
* is finer than the underlying fs, go check to see if
* we must zero out the start of this block.
*/
if (unlikely(dio->blkfactor && !dio->start_zero_done))
dio_zero_block(dio, 0);
/*
* Work out, in this_chunk_blocks, how much disk we
* can add to this page
*/
this_chunk_blocks = dio->blocks_available;
u = (PAGE_SIZE - offset_in_page) >> blkbits;
if (this_chunk_blocks > u)
this_chunk_blocks = u;
u = dio->final_block_in_request - dio->block_in_file;
if (this_chunk_blocks > u)
this_chunk_blocks = u;
this_chunk_bytes = this_chunk_blocks << blkbits;
BUG_ON(this_chunk_bytes == 0);
dio->boundary = buffer_boundary(map_bh);
ret = submit_page_section(dio, page, offset_in_page,
this_chunk_bytes, dio->next_block_for_io);
if (ret) {
page_cache_release(page);
goto out;
}
dio->next_block_for_io += this_chunk_blocks;
dio->block_in_file += this_chunk_blocks;
block_in_page += this_chunk_blocks;
dio->blocks_available -= this_chunk_blocks;
next_block:
BUG_ON(dio->block_in_file > dio->final_block_in_request);
if (dio->block_in_file == dio->final_block_in_request)
break;
}
/* Drop the ref which was taken in get_user_pages() */
page_cache_release(page);
block_in_page = 0;
}
out:
return ret;
}
/*
* Releases both i_mutex and i_alloc_sem
*/
static ssize_t
direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
const struct iovec *iov, loff_t offset, unsigned long nr_segs,
unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
struct dio *dio)
{
unsigned long user_addr;
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
unsigned long flags;
int seg;
ssize_t ret = 0;
ssize_t ret2;
size_t bytes;
dio->inode = inode;
dio->rw = rw;
dio->blkbits = blkbits;
dio->blkfactor = inode->i_blkbits - blkbits;
dio->block_in_file = offset >> blkbits;
dio->get_block = get_block;
dio->end_io = end_io;
dio->final_block_in_bio = -1;
dio->next_block_for_io = -1;
dio->iocb = iocb;
dio->i_size = i_size_read(inode);
spin_lock_init(&dio->bio_lock);
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
dio->refcount = 1;
/*
* In case of non-aligned buffers, we may need 2 more
* pages since we need to zero out first and last block.
*/
if (unlikely(dio->blkfactor))
dio->pages_in_io = 2;
for (seg = 0; seg < nr_segs; seg++) {
user_addr = (unsigned long)iov[seg].iov_base;
dio->pages_in_io +=
((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
- user_addr/PAGE_SIZE);
}
for (seg = 0; seg < nr_segs; seg++) {
user_addr = (unsigned long)iov[seg].iov_base;
dio->size += bytes = iov[seg].iov_len;
/* Index into the first page of the first block */
dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
dio->final_block_in_request = dio->block_in_file +
(bytes >> blkbits);
/* Page fetching state */
dio->head = 0;
dio->tail = 0;
dio->curr_page = 0;
dio->total_pages = 0;
if (user_addr & (PAGE_SIZE-1)) {
dio->total_pages++;
bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
}
dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
dio->curr_user_address = user_addr;
ret = do_direct_IO(dio);
dio->result += iov[seg].iov_len -
((dio->final_block_in_request - dio->block_in_file) <<
blkbits);
if (ret) {
dio_cleanup(dio);
break;
}
} /* end iovec loop */
if (ret == -ENOTBLK && (rw & WRITE)) {
/*
* The remaining part of the request will be
* be handled by buffered I/O when we return
*/
ret = 0;
}
/*
* There may be some unwritten disk at the end of a part-written
* fs-block-sized block. Go zero that now.
*/
dio_zero_block(dio, 1);
if (dio->cur_page) {
ret2 = dio_send_cur_page(dio);
if (ret == 0)
ret = ret2;
page_cache_release(dio->cur_page);
dio->cur_page = NULL;
}
if (dio->bio)
dio_bio_submit(dio);
/* All IO is now issued, send it on its way */
blk_run_address_space(inode->i_mapping);
/*
* It is possible that, we return short IO due to end of file.
* In that case, we need to release all the pages we got hold on.
*/
dio_cleanup(dio);
/*
* All block lookups have been performed. For READ requests
* we can let i_mutex go now that its achieved its purpose
* of protecting us from looking up uninitialized blocks.
*/
if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
mutex_unlock(&dio->inode->i_mutex);
/*
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
* The only time we want to leave bios in flight is when a successful
* partial aio read or full aio write have been setup. In that case
* bio completion will call aio_complete. The only time it's safe to
* call aio_complete is when we return -EIOCBQUEUED, so we key on that.
* This had *better* be the only place that raises -EIOCBQUEUED.
*/
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
BUG_ON(ret == -EIOCBQUEUED);
if (dio->is_async && ret == 0 && dio->result &&
((rw & READ) || (dio->result == dio->size)))
ret = -EIOCBQUEUED;
[PATCH] dio: formalize bio counters as a dio reference count Previously we had two confusing counts of bio progress. 'bio_count' was decremented as bios were processed and freed by the dio core. It was used to indicate final completion of the dio operation. 'bios_in_flight' reflected how many bios were between submit_bio() and bio->end_io. It was used by the sync path to decide when to wake up and finish completing bios and was ignored by the async path. This patch collapses the two notions into one notion of a dio reference count. bios hold a dio reference when they're between submit_bio and bio->end_io. Since bios_in_flight was only used in the sync path it is now equivalent to dio->refcount - 1 which accounts for direct_io_worker() holding a reference for the duration of the operation. dio_bio_complete() -> finished_one_bio() was called from the sync path after finding bios on the list that the bio->end_io function had deposited. finished_one_bio() can not drop the dio reference on behalf of these bios now because bio->end_io already has. The is_async test in finished_one_bio() meant that it never actually did anything other than drop the bio_count for sync callers. So we remove its refcount decrement, don't call it from dio_bio_complete(), and hoist its call up into the async dio_bio_complete() caller after an explicit refcount decrement. It is renamed dio_complete_aio() to reflect the remaining work it actually does. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:59 +08:00
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
if (ret != -EIOCBQUEUED)
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
dio_await_completion(dio);
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
/*
* Sync will always be dropping the final ref and completing the
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
* operation. AIO can if it was a broken operation described above or
* in fact if all the bios race to complete before we get here. In
* that case dio_complete() translates the EIOCBQUEUED into the proper
* return code that the caller will hand to aio_complete().
*
* This is managed by the bio_lock instead of being an atomic_t so that
* completion paths can drop their ref and use the remaining count to
* decide to wake the submission path atomically.
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
*/
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
spin_lock_irqsave(&dio->bio_lock, flags);
ret2 = --dio->refcount;
spin_unlock_irqrestore(&dio->bio_lock, flags);
[PATCH] dio: lock refcount operations The wait_for_more_bios() function name was poorly chosen. While looking to clean it up it I noticed that the dio struct refcounting between the bio completion and dio submission paths was racey. The bio submission path was simply freeing the dio struct if atomic_dec_and_test() indicated that it dropped the final reference. The aio bio completion path was dereferencing its dio struct pointer *after dropping its reference* based on the remaining number of references. These two paths could race and result in the aio bio completion path dereferencing a freed dio, though this was not observed in the wild. This moves the refcount under the bio lock so that bio completion can drop its reference and decide to wake all in one atomic step. Once testing and waking is locked dio_await_one() can test its sleeping condition and mark itself uninterruptible under the lock. It gets simpler and wait_for_more_bios() disappears. The addition of the interrupt masking spin lock acquiry in dio_bio_submit() looks alarming. This lock acquiry existed in that path before the recent dio completion patch set. We shouldn't expect significant performance regression from returning to the behaviour that existed before the completion clean up work. This passed 4k block ext3 O_DIRECT fsx and aio-stress on an SMP machine. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:07 +08:00
if (ret2 == 0) {
[PATCH] dio: centralize completion in dio_complete() There have been a lot of bugs recently due to the way direct_io_worker() tries to decide how to finish direct IO operations. In the worst examples it has failed to call aio_complete() at all (hang) or called it too many times (oops). This set of patches cleans up the completion phase with the goal of removing the complexity that lead to these bugs. We end up with one path that calculates the result of the operation after all off the bios have completed. We decide when to generate a result of the operation using that path based on the final release of a refcount on the dio structure. I tried to progress towards the final state in steps that were relatively easy to understand. Each step should compile but I only tested the final result of having all the patches applied. I've tested these on low end PC drives with aio-stress, the direct IO tests I could manage to get running in LTP, orasim, and some home-brew functional tests. In http://lkml.org/lkml/2006/9/21/103 IBM reports success with ext2 and ext3 running DIO LTP tests. They found that XFS bug which has since been addressed in the patch series. This patch: The mechanics which decide the result of a direct IO operation were duplicated in the sync and async paths. The async path didn't check page_errors which can manifest as silently returning success when the final pointer in an operation faults and its matching file region is filled with zeros. The sync path and async path differed in whether they passed errors to the caller's dio->end_io operation. The async path was passing errors to it which trips an assertion in XFS, though it is apparently harmless. This centralizes the completion phase of dio ops in one place. AIO will now return EFAULT consistently and all paths fall back to the previously sync behaviour of passing the number of bytes 'transferred' to the dio->end_io callback, regardless of errors. dio_await_completion() doesn't have to propogate EIO from non-uptodate bios now that it's being propogated through dio_complete() via dio->io_error. This lets it return void which simplifies its sole caller. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:20:54 +08:00
ret = dio_complete(dio, offset, ret);
[PATCH] dio: only call aio_complete() after returning -EIOCBQUEUED The only time it is safe to call aio_complete() is when the ->ki_retry function returns -EIOCBQUEUED to the AIO core. direct_io_worker() has historically done this by relying on its caller to translate positive return codes into -EIOCBQUEUED for the aio case. It did this by trying to keep conditionals in sync. direct_io_worker() knew when finished_one_bio() was going to call aio_complete(). It would reverse the test and wait and free the dio in the cases it thought that finished_one_bio() wasn't going to. Not surprisingly, it ended up getting it wrong. 'ret' could be a negative errno from the submission path but it failed to communicate this to finished_one_bio(). direct_io_worker() would return < 0, it's callers wouldn't raise -EIOCBQUEUED, and aio_complete() would be called. In the future finished_one_bio()'s tests wouldn't reflect this and aio_complete() would be called for a second time which can manifest as an oops. The previous cleanups have whittled the sync and async completion paths down to the point where we can collapse them and clearly reassert the invariant that we must only call aio_complete() after returning -EIOCBQUEUED. direct_io_worker() will only return -EIOCBQUEUED when it is not the last to drop the dio refcount and the aio bio completion path will only call aio_complete() when it is the last to drop the dio refcount. direct_io_worker() can ensure that it is the last to drop the reference count by waiting for bios to drain. It does this for sync ops, of course, and for partial dio writes that must fall back to buffered and for aio ops that saw errors during submission. This means that operations that end up waiting, even if they were issued as aio ops, will not call aio_complete() from dio. Instead we return the return code of the operation and let the aio core call aio_complete(). This is purposely done to fix a bug where AIO DIO file extensions would call aio_complete() before their callers have a chance to update i_size. Now that direct_io_worker() is explicitly returning -EIOCBQUEUED its callers no longer have to translate for it. XFS needs to be careful not to free resources that will be used during AIO completion if -EIOCBQUEUED is returned. We maintain the previous behaviour of trying to write fs metadata for O_SYNC aio+dio writes. Signed-off-by: Zach Brown <zach.brown@oracle.com> Cc: Badari Pulavarty <pbadari@us.ibm.com> Cc: Suparna Bhattacharya <suparna@in.ibm.com> Acked-by: Jeff Moyer <jmoyer@redhat.com> Cc: <xfs-masters@oss.sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-12-10 18:21:05 +08:00
kfree(dio);
} else
BUG_ON(ret != -EIOCBQUEUED);
return ret;
}
/*
* This is a library function for use by filesystem drivers.
* The locking rules are governed by the dio_lock_type parameter.
*
* DIO_NO_LOCKING (no locking, for raw block device access)
* For writes, i_mutex is not held on entry; it is never taken.
*
* DIO_LOCKING (simple locking for regular files)
* For writes we are called under i_mutex and return with i_mutex held, even
* though it is internally dropped.
* For reads, i_mutex is not held on entry, but it is taken and dropped before
* returning.
*
* DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
* uninitialised data, allowing parallel direct readers and writers)
* For writes we are called without i_mutex, return without it, never touch it.
* For reads we are called under i_mutex and return with i_mutex held, even
* though it may be internally dropped.
*
* Additional i_alloc_sem locking requirements described inline below.
*/
ssize_t
__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
struct block_device *bdev, const struct iovec *iov, loff_t offset,
unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
int dio_lock_type)
{
int seg;
size_t size;
unsigned long addr;
unsigned blkbits = inode->i_blkbits;
unsigned bdev_blkbits = 0;
unsigned blocksize_mask = (1 << blkbits) - 1;
ssize_t retval = -EINVAL;
loff_t end = offset;
struct dio *dio;
int release_i_mutex = 0;
int acquire_i_mutex = 0;
if (rw & WRITE)
rw = WRITE_SYNC;
if (bdev)
bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
if (offset & blocksize_mask) {
if (bdev)
blkbits = bdev_blkbits;
blocksize_mask = (1 << blkbits) - 1;
if (offset & blocksize_mask)
goto out;
}
/* Check the memory alignment. Blocks cannot straddle pages */
for (seg = 0; seg < nr_segs; seg++) {
addr = (unsigned long)iov[seg].iov_base;
size = iov[seg].iov_len;
end += size;
if ((addr & blocksize_mask) || (size & blocksize_mask)) {
if (bdev)
blkbits = bdev_blkbits;
blocksize_mask = (1 << blkbits) - 1;
if ((addr & blocksize_mask) || (size & blocksize_mask))
goto out;
}
}
dio = kzalloc(sizeof(*dio), GFP_KERNEL);
retval = -ENOMEM;
if (!dio)
goto out;
/*
* For block device access DIO_NO_LOCKING is used,
* neither readers nor writers do any locking at all
* For regular files using DIO_LOCKING,
* readers need to grab i_mutex and i_alloc_sem
* writers need to grab i_alloc_sem only (i_mutex is already held)
* For regular files using DIO_OWN_LOCKING,
* neither readers nor writers take any locks here
*/
dio->lock_type = dio_lock_type;
if (dio_lock_type != DIO_NO_LOCKING) {
/* watch out for a 0 len io from a tricksy fs */
if (rw == READ && end > offset) {
struct address_space *mapping;
mapping = iocb->ki_filp->f_mapping;
if (dio_lock_type != DIO_OWN_LOCKING) {
mutex_lock(&inode->i_mutex);
release_i_mutex = 1;
}
retval = filemap_write_and_wait_range(mapping, offset,
end - 1);
if (retval) {
kfree(dio);
goto out;
}
if (dio_lock_type == DIO_OWN_LOCKING) {
mutex_unlock(&inode->i_mutex);
acquire_i_mutex = 1;
}
}
if (dio_lock_type == DIO_LOCKING)
/* lockdep: not the owner will release it */
down_read_non_owner(&inode->i_alloc_sem);
}
/*
* For file extending writes updating i_size before data
* writeouts complete can expose uninitialized blocks. So
* even for AIO, we need to wait for i/o to complete before
* returning in this case.
*/
dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
(end > i_size_read(inode)));
retval = direct_io_worker(rw, iocb, inode, iov, offset,
nr_segs, blkbits, get_block, end_io, dio);
if (rw == READ && dio_lock_type == DIO_LOCKING)
release_i_mutex = 0;
out:
if (release_i_mutex)
mutex_unlock(&inode->i_mutex);
else if (acquire_i_mutex)
mutex_lock(&inode->i_mutex);
return retval;
}
EXPORT_SYMBOL(__blockdev_direct_IO);